APIR
APIR · Trust Center

Check us the way
we ask you to check anyone.

Most trust pages are a list of promises and a badge. Every line on this one carries a mark: checkable, meaning it links to something you can verify without trusting us; our word, meaning it is a statement about our own systems; or not held.

Checkable Our word Not held

Ed25519

Every passport signed

RFC 3161

Outside timestamps, nightly

TLS 1.2+

In transit, 1.3 offered

None held

SOC 2, ISO 27001, pen test

Last updated 28 September 2026

01 · Certifications

What we hold: nothing, yet

A small company should say this plainly. When an audit is engaged, this section will name the auditor and the date. Until then, the rest of this page is built so you need neither.

Not held

SOC 2 Type II

Not held. No auditor is engaged. Our database and hosting providers hold their own SOC 2 reports; those are theirs, not ours.

Not held

ISO/IEC 27001

Not held. No registrar is engaged.

Not held

Penetration test

Not done by a third party. If you find something, the disclosure section below is how to tell us.

Not held

ISO/IEC 42001

Not held. We do not claim alignment to a standard nobody has assessed us against.

02 · Check it yourself

What you can verify without us

The part no badge can give you. Each of these opens the real thing, and none of them asks you to trust APIR.

The public signing keys

Checkable

Every Trust Passport is signed with Ed25519. The public keys are published; once you hold them, checking a public passport's signature needs no call to us. The signed bytes are set out in the specification.

A passport, with its chain

Checkable

The Originator’s papers: the signature, and the Black Box chain behind it recomputed from stored content, as of the reading shown on the page.

An outside timestamp

Checkable

Seals and the Black Box sit on hash chains whose heads are timestamped each night by an RFC 3161 authority we do not control. The specimen seal’s portable proof carries the token; openssl checks it against the authority’s certificate.

A checker that runs unplugged

Checkable

One page, no network requests allowed by its own security policy. Drop a proof and a file into it and check a seal with APIR switched off. apir-record, MIT-licensed Python, checks an owner’s export and a public agent’s stamp on your own machine.

Every public passport

Checkable

The radar re-verifies each public passport’s signature against the published key on every reading.

Encryption in transit

Checkable

TLS 1.2 is the minimum on apir.ai and 1.3 is offered. Anyone can confirm it: openssl s_client -connect apir.ai:443.

03 · Security controls

How the platform is built

Statements about our own systems. They are true as of the date on this page, and they are ours: a reviewer who needs more can ask, and a person answers.

Row-level security on every table

Our word

Row-level security is switched on for every table in the database, and none is exempt. Measured 28 September 2026.

Signing keys that never leave the vault

Our word

Private signing keys are generated inside the edge runtime, sealed in Supabase Vault, and never exported.

Encrypted at rest

Our word

AES-256, applied by our database provider to the storage underneath.

No secret in the browser

Our word

No secret in the repository or the client bundle; the only key a browser sees is the public one it needs. The service-role key stays on the server.

One administrator

Our word

Platform administration is one fixed identity, gated on the server as well as hidden in the interface. There is no role that grants it to anyone else.

Two-factor sign-in

Our word

TOTP two-factor is available in account settings.

Rate limits on the free checks

Our word

The free, no-account checks are limited per caller. On since 28 September 2026.

Gated change

Our word

Production deploys only from a build that passes its typecheck. Lint and tests run on every push.

04 · AI and your data

What happens to what you give us

We train no models

Our word

APIR trains no model, on your data or anyone’s. It calls models through their providers’ APIs.

Where a model call goes

Our word

To Anthropic directly first. If that fails, through OpenRouter to another provider, with an open-weight model as the last resort. Web-grounded search and some bulk work go to Google’s Gemini API.

An agent stages; a person sends

Our word

No agent emails, posts, signs, lodges or charges on its own. A database trigger refuses any draft marked approved or sent with no human reviewer recorded, whoever the caller.

Your file stays with you

Checkable

Checking a file against the seal record, and the Numbers Check, fingerprint the file in your browser. The bytes are never sent. Watch the network tab.

05 · Subprocessors

Who else touches your data

Every service that processes data for APIR, what it does, and where. Current as of 28 September 2026.

SubprocessorPurpose
SupabaseDatabase, auth, storage, edge functions
CloudflareHosting, CDN, edge runtime, rate limits
StripePayments and billing
AnthropicModel calls (Claude)
OpenRouterFallback model routing to other providers
GoogleGemini API (search, bulk work), Places (business lookups)
ResendEmail
PostHogProduct analytics
freeTSARFC 3161 timestamps; receives a hash, never content
06 · Retention and disclosure

Keeping, deleting, reporting

Our word

Where it lives

The primary database is in Singapore (ap-southeast-1). Subprocessors above operate where the table says.

Our word

What is deleted, and what is not

Account data is deleted on a verified request. The Black Box and the seal chain are append-only by design and are not erased: a record that could be quietly removed is one nobody could rely on. Think before you seal.

The Data Processing Agreement
Our word

Found something?

Email ripa@apir.ai. We acknowledge within 48 hours and coordinate disclosure within 90 days. Also at /.well-known/security.txt.

Procurement

Have a security questionnaire?

Send it. A person answers every line, and anything we do not hold is written as not held.

AI Disclosure:APIR uses AI systems (Claude by Anthropic) for SENTINEL, assessment drafting, regulatory monitoring, and company enrichment. AI-generated outputs are labelled as such.

Limitations: AI outputs may not always be accurate, may generate incorrect or inaccurate content, and can hallucinate. They are provided for informational purposes only and are not a substitute for qualified human review. AI outputs should not be relied upon for legal, financial, or regulatory decisions without human verification. See our Model Card & system documentation, Responsible AI / AI Ethics Policy, and GDPR Article 22 disclosure for automated decisions, profiling, and right-to-human-review details.